I’m still amazed at Claude Code’s performance in my codebase. Within days, I rewrote what took weeks before. Here’s what I learned in the last days:
CLAUDE.md’s are important
Where the README.md is targeted at humans getting up to speed with a project, the CLAUDE.md is the same for coding agents. This is my starting point, which tells Claude to update the Claude.md. After all, he’s the agent, so he should know best what the next agent needs. And it’s one less dull task for me.
I like that you can have multiple CLAUDE.md files in your project. In my testing folder, there’s a CLAUDE.md that describes everything relevant for testing. This avoids cluttering the central file with irrelevant stuff. Claude automatically reads the files before others in the folder.
Agentic coding just solves the coding part
If you’ve worked on a real software project, you know coding is maybe 10-20% of the work. Here’s a non-complete list of things that come to mind:
- Aligning with the customer or product owner on the product’s goals.
- exploring and communicating what’s technically possible and viable.
- Based on this, agreeing on the project scope.
- Slicing the work into the right packages
- Setting up infrastructure
- Testing
- Verifying with customers it solves their problem
Still, agentic coding is great: It will enable more people to experience these problems themselves, not just hearing it from their developers. Imagine your product owner can build the first prototypes and develop a sense of what’s possible!
Slicing the work is important
Developers have a huge advantage over non-developers in this new world of agentic coding because they know how to slice the work. The planning mode of Claude is nice, but the results from the coding assistant depend on the right size of work. Pick a too-small task and you’ll be disappointed. The overhead of testing and committing is real. Pick a task too big and Claude will be overloaded and won’t find a good solution. Just like with real developers.
How Does Claude Code Work?
Last week I promised to answer this question. Unfortunately, Claude Code’s source code isn’t available. Here are my observations:
Anthropic has the best coding models right now.
Anthropic’s Claude models are ahead of the competition. Other companies like Meta switched to Claude instead of their own models. This is definitely not an easy step for Meta, given their investment in Llama and not done for small improvements.
Claude Code works well due to its good model. With my Max plan, I’m using only the newest Opus model and it’s a big step up from the already good older Sonnet model.
Claude can split the work into small chunks
There’s an interesting tool called claude-trace that lets you see ClaudeCode’s API requests. Simon Willison covered this in a blog post, and I liked this video covering it.
The agent breaks down each task into smaller ones, waits for the result, and puts it back together. Many requests can be done in parallel. This is also why Claude’s code is 4-5 times more expensive than just using Cursor, for example. It makes a massive amount of requests. Definitely not something you can run with local LLMs on your machine at the moment.
Claude can self-verify its solution
Yesterday, Claude changed my code and broke the web service, causing localhost:3000 to load forever. What’s cool is that Claude can take over the command line, make a change, restart the server, and test it. After observing it for a week, I’m not sure Claude is much smarter than Cursor. It’s just much more autonomous.
Conclusion
My assumption for now is: The more your codebase allows for self-verification and instant feedback, whether this change is good or bad, the more you will get out of coding agents.
We’re close to a future where companies with a very good stack will allow their coding agents to develop a small feature behind a feature flag, ship it live to a small percentage of users, monitor performance and user impact, and report back. I assume the biggest problem for most companies will not be the coding agent, but getting your codebase to a state that enables this.
What to Print this Week
This newsletter started out on 3D printing. If you haven't had any contact with it, you should, it's great! Here's the most interesting and funniest projects I saw last week.
This is my favourite in the Diorama contest: A cute mini fire escape stairway, attachable to your wall.
|
|
Mini Fire Escape Diorama
|
Lamps continue to be a hot topic, and they get more beautiful every month.
|
|
Lampshade - Triangle Displaced
|
Ironing is an advanced setting to get more perfect surfaces, this is an easy print to experience the difference.
|
|
Hi 👋, I'm Stefan!
This is my weekly newsletter about new technology hypes in general and AI in specific. Feel free to forward this mail to people who should read it. If this mail was forwarded to you, please subscribe here.
|
|
